How sml-compbio/oncomark works
Main Components
• Core Library (OncoMark Package): This includes files such as oncomark.py, hallmark_feature.txt, hallmark_model.keras, and hallmark_scaler.joblib within the OncoMark directory. This is the central component implementing the deep learning prediction logic.
• Data Preprocessing & Model Training: Under the src directory you have several notebooks (e.g., Model_Train_Data_Prep.ipynb, Model_Train_and_Int_Val.ipynb, etc.) which indicate that there is a data preparation and training pipeline. These components reflect the model training, validation, and data handling stages.
• Web Server Interface: The webserver directory contains its own copy of key model assets (e.g., OncoMark.py, mlm.py, ulm.py, plotting.py) plus sample data (collectri_df.csv, progeny_df.csv) and a README for the web server. This component is responsible for exposing the OncoMark functionality via a web interface.
• Testing/Validation: The test directory contains test notebooks and sample test data to ensure model accuracy and integration.
• Documentation & Deployment Artifacts: The docs folder, Workflow.md, and configuration files (mkdocs.yml, readthedocs.yaml, setup.py, requirements.txt) form the supportive components necessary for documentation and deployment.
Detail the Relationships and Data Flows
• Data Flow: Start with the user-provided transcriptomics data (from tumor biopsies) entering the system. In the Python API, data is processed, passed to the ML model (loaded from keras and scaled using joblib), and then predictions are returned.
• Web Server Interaction: For the web interface, show how the web server accepts data requests, passes them to the core prediction library, and returns predictions for visualization or further analysis.
• Testing: Incorporate the model testing pipeline showing how automated tests (with notebooks and CSV sample data in the test folder) validate the outputs.
Architectural Patterns and Technologies
• Multi-layer Architecture: Represent a clear separation of concerns:
• Input Layer – where data is ingested.
• Business Logic Layer – the core prediction functions (the OncoMark API and model components).
• Presentation Layer – the web server component serving a user interface.
• That the project uses a multi-task learning (deep learning) approach as part of its neural network design. The core components (using keras for the model and joblib for scaling) are central to this design.
• Use of notebooks for exploratory data analysis, model training, and validation. These act as a developmental and experimental interface.
• Integration of external libraries/frameworks (such as keras, joblib, and any web server framework employed in the webserver folder) as well as configuration for documentation deployment (mkdocs, readthedocs).